Higher-Order Control Flow

One of the primary roles of the standard library is to make it easier to express high-level algorithmic ideas as quantum programs. Thus, the Q# canon provides a variety of different flow control constructs, each implemented using partial application of functions and operations. Jumping immediately into an example, consider the case in which one wants to construct a "CNOT ladder" on a register:

let nQubits = Length(register);
CNOT(register[0], register[1]);
CNOT(register[1], register[2]);
// ...
CNOT(register[nQubits - 2], register[nQubits - 1]);

We can express this pattern by using iteration and for loops:

for idxQubit in 0..nQubits - 2 {
    CNOT(register[idxQubit], register[idxQubit + 1]);
}

Expressed in terms of ApplyToEachCA operation and array manipulation functions such as Zipped function, however, this is much shorter and easier to read:

ApplyToEachCA(CNOT, Zip(register[0..nQubits - 2], register[1..nQubits - 1]));

In the rest of this section, we will provide a number of examples of how to use the various flow control operations and functions provided by the canon to compactly express quantum programs.

Applying Operations and Functions over Arrays and Ranges

One of the primary abstractions provided by the canon is that of iteration. For instance, consider a unitary of the form $U \otimes U \otimes \cdots \otimes U$ for a single-qubit unitary $U$. In Q#, we might use IndexRange function to represent this as as a for loop over a register:

/// # Summary
/// Applies $H$ to all qubits in a register.
operation ApplyHadamardToAll(
    register : Qubit[])
: Unit is Adj + Ctl {
    for qubit in register {
        H(qubit);
    }
}

We can then use this new operation as HAll(register) to apply $H \otimes H \otimes \cdots \otimes H$.

This is cumbersome to do, however, especially in a larger algorithm. Moreover, this approach is specialized to the particular operation that we wish to apply to each qubit. We can use the fact that operations are first-class to express this algorithmic concept more explicitly:

ApplyToEachCA(H, register);

Here, the suffix CA indicates that the call to ApplyToEach is itself adjointable and controllable. Thus, if U supports Adjoint and Controlled, the following lines are equivalent:

Adjoint ApplyToEachCA(U, register);
ApplyToEachCA(Adjoint U, register);

In particular, this means that calls to ApplyToEachCA can appear in operations for which an adjoint specialization is auto-generated. Similarly, ApplyToEachIndex operation is useful for representing patterns of the form U(0, targets[0]); U(1, targets[1]); ..., and offers versions for each combination of functors supported by its input.

Tip

ApplyToEach is type-parameterized such that it can be used with operations that take inputs other than Qubit. For example, suppose that codeBlocks is an array of LogicalRegister user defined type values that need to be recovered. Then ApplyToEach(Recover(code, recoveryFn, _), codeBlocks) will apply the error-correcting code code and recovery function recoveryFn to each block independently. This holds even for classical inputs: ApplyToEach(R(_, _, qubit), [(PauliX, PI() / 2.0); (PauliY(), PI() / 3.0])) will apply a rotation of $\pi / 2$ about $X$ followed by a rotation of $pi / 3$ about $Y$.

The Q# canon also provides support for classical enumeration patterns familiar to functional programming. For instance, Fold function implements the pattern $f(f(f(s_{\text{initial}}, x_0), x_1), \dots)$ for reducing a function over a list. This pattern can be used to implement sums, products, minima, maxima and other such functions:

open Microsoft.Quantum.Arrays;
function Plus(a : Int, b : Int) : Int { return a + b; }
function Sum(xs : Int[]) {
    return Fold(Sum, 0, xs);
}

Similarly, functions like Mapped function and MappedByIndex function can be used to express functional programming concepts in Q#.

Composing Operations and Functions

The control flow constructs offered by the canon take operations and functions as their inputs, such that it is helpful to be able to compose several operations or functions into a single callable. For instance, the pattern $UVU^{\dagger}$ is extremely common in quantum programming, such that the canon provides the operation ApplyWith operation as an abstraction for this pattern. This abstraction also allows for more efficient compliation into circuits, as Controlled acting on the sequence U(qubit); V(qubit); Adjoint U(qubit); does not need to act on each U. To see this, let $c(U)$ be the unitary representing Controlled U([control], target) and let $c(V)$ be defined in the same way. Then for an arbitrary state $\ket{\psi}$, \begin{align} c(U) c(V) c(U)^\dagger \ket{1} \otimes \ket{\psi} & = \ket{1} \otimes (UVU^{\dagger} \ket{\psi}) \\ & = (\boldone \otimes U) (c(V)) (\boldone \otimes U^\dagger) \ket{1} \otimes \ket{\psi}. \end{align} by the definition of Controlled. On the other hand, \begin{align} c(U) c(V) c(U)^\dagger \ket{0} \otimes \ket{\psi} & = \ket{0} \otimes \ket{\psi} \\ & = \ket{0} \otimes (UU^\dagger \ket{\psi}) \\ & = (\boldone \otimes U) (c(V)) (\boldone \otimes U^\dagger) \ket{0} \otimes \ket{\psi}. \end{align} By linearity, we can conclude that we can factor $U$ out in this way for all input states. That is, $c(UVU^\dagger) = U c(V) U^\dagger$. Since controlling operations can be expensive in general, using controlled variants such as WithC and WithCA can help reduce the number of control functors that need to be applied.

Note

One other consequence of factoring out $U$ is that we need not even know how to apply the Controlled functor to U. ApplyWithCA therefore has a weaker signature than might be expected:

ApplyWithCA<'T> : (('T => Unit is Adj),
    ('T => Unit is Adj + Ctl), 'T) => Unit

Similarly, Bound function produces operations which apply a sequence of other operations in turn. For instance, the following are equivalent:

H(qubit); X(qubit);
Bound([H, X], qubit);

Combining with iteration patterns can make this especially useful:

// Bracket the quantum Fourier transform with $XH$ on each qubit.
ApplyWith(ApplyToEach(Bound([H, X]), _), QFT, _);

Time-Ordered Composition

We can go still further by thinking of flow control in terms of partial application and classical functions, and can model even fairly sophisticated quantum concepts in terms of classical flow control. This analogy is made precise by the recognition that unitary operators correspond exactly to the side effects of calling operations, such that any decomposition of unitary operators in terms of other unitary operators corresponds to constructing a particular calling sequence for classical subroutines which emit instructions to act as particular unitary operators. Under this view, Bound is precisely the representation of the matrix product, since Bound([A, B])(target) is equivalent to A(target); B(target);, which in turn is the calling sequence corresponding to $BA$.

A more sophisticated example is the Trotter–Suzuki expansion. As discussed in the Dynamical Generator Representation section of data structures, the Trotter–Suzuki expansion provides a particularly useful way of expressing matrix exponentials. For instance, applying the expansion at its lowest order yields that for any operators $A$ and $B$ such that $A = A^\dagger$ and $B = B^\dagger$, \begin{align} \tag{★} \label{eq:trotter-suzuki-0} \exp(i [A + B] t) = \lim_{n\to\infty} \left(\exp(i A t / n) \exp(i B t / n)\right)^n. \end{align} Colloquially, this says that we can approximately evolve a state under $A + B$ by alternately evolving under $A$ and $B$ alone. If we represent evolution under $A$ by an operation A : (Double, Qubit[]) => Unit that applies $e^{i t A}$, then the representation of the Trotter–Suzuki expansion in terms of rearranging calling sequences becomes clear. Concretely, given an operation U : ((Int, Double, Qubit[]) => Unit is Adj + Ctl such that A = U(0, _, _) and B = U(1, _, _), we can define a new operation representing the integral of U at time $t$ by generating sequences of the form

U(0, time / Float(nSteps), target);
U(1, time / Float(nSteps), target);
U(0, time / Float(nSteps), target);
U(1, time / Float(nSteps), target);
// ...

At this point, we can now reason about the Trotter–Suzuki expansion without reference to quantum mechanics at all. The expansion is effectively a very particular iteration pattern motivated by $\eqref{eq:trotter-suzuki-0}$. This iteration pattern is implemented by DecomposedIntoTimeStepsCA function:

// The 2 indicates how many terms we need to decompose,
// while the 1 indicates that we are using the
// first-order Trotter–Suzuki decomposoition.
DecomposeIntoTimeStepsCA((2, U), 1);

The signature of DecomposeIntoTimeStepsCA follows a common pattern in Q#, where collections that may be backed either by arrays or by something which compute elements on the fly are represented by tuples whose first elements are Int values indicating their lengths.

Putting it Together: Controlling Operations

Finally, the canon builds on the Controlled functor by providing additional ways to condition quantum operations. It is common, especially in quantum arithmetic, to condition operations on computational basis states other than $\ket{0\cdots 0}$. Using the control operations and functions introduced above, we can more general quantum conditions in a single statement. Let's jump in to how ControlledOnBitString function does it (sans type parameters), then we'll break down the pieces one by one. The first thing we'll need to do is to define an operation which actually does the heavy lifting of implementing the control on arbitrary computational basis states. We won't call this operation directly, however, and so we add _ to the beginning of the name to indicate that it's an implementation of another construct elsewhere.

operation _ControlledOnBitString(
    bits : Bool[],
    oracle: (Qubit[] => Unit is Adj + Ctl),
    controlRegister : Qubit[],
    targetRegister: Qubit[])
: Unit is Adj + Ctl

Note that we take a string of bits, represented as a Bool array, that we use to specify the conditioning that we want to apply to the operation oracle that we are given. Since this operation actually does the application directly, we also need to take the control and target registers, both represented here as Qubit[].

Next, we note that controlling on the computational basis state $\ket{\vec{s}} = \ket{s_0 s_1 \dots s_{n - 1}}$ for a bit string $\vec{s}$ can be realized by transforming $\ket{\vec{s}}$ into $\ket{0\cdots 0}$. In particular, $\ket{\vec{s}} = X^{s_0} \otimes X^{s_1} \otimes \cdots \otimes X^{s_{n - 1}} \ket{0\cdots 0}$. Since $X^{\dagger} = X$, this implies that $\ket{0\dots 0} = X^{s_0} \otimes X^{s_1} \otimes \cdots \otimes X^{s_{n - 1}} \ket{\vec{s}}$. Thus, we can apply $P = X^{s_0} \otimes X^{s_1} \otimes \cdots \otimes X^{s_{n - 1}}$, apply Controlled oracle, and transform back to $\vec{s}$. This construction is precisely ApplyWith, so we write the body of our new operation accordingly:

{
    ApplyWithCA(
        ApplyPauliFromBitString(PauliX, false, bits, _),
        (Controlled oracle)(_, targetRegister),
        controlRegister
    );
}

Here, we have used ApplyPauliFromBitString operation to apply $P$, partially applying over its target for use with ApplyWith. Note, however, that we need to transform the control register to our desired form, so we partially apply the inner operation (Controlled oracle) on the target. This leaves ApplyWith acting to bracket the control register with $P$, exactly as we desired.

At this point, we could be done, but it is somehow unsatisfying that our new operation does not "feel" like applying the Controlled functor. Thus, we finish defining our new control flow concept by writing a function that takes the oracle to be controlled and that returns a new operation. In this way, our new function looks and feels very much like Controlled, illustrating that we can easily define powerful new control flow constructs using Q# and the canon together:

function ControlledOnBitString(
    bits : Bool[],
    oracle: (Qubit[] => Unit is Adj + Ctl))
: ((Qubit[], Qubit[]) => Unit is Adj + Ctl) {
    return _ControlledOnBitString(bits, oracle, _, _);
}